This text is an experiment. So far, this blog has been 100% human written (the quality of my typos can attests to that). This time, I wanted to test something new. I've come up with the ideas, had the experiences, and everything a writer should have before writing, but instead of writing like I normally would, I yapped into my dictation tool for 20 minutes following a rough outline, then asked AI to clean it up for me. I edited the result myself. I've added my review of this experiment at the end.
The first tool I made for myself with AI was a Substack feed blocker.
I enjoy Substack. I think it's a genuinely good medium. I'm not always sure I enjoy the culture (especially on the French side of things) but I enjoy the medium. I like newsletters/blogs. I like having a little website where I can publish my thoughts and send them directly to people. I like Substack Notes. There are plenty of cool people to follow there.
But the feed was regularly making me mad. I would open Substack to publish something, check a reply or read someone I liked, and I'd end up leaving with this vague sense of irritation about the world.
Irritation can be useful fuel for action. This wasn't that kind. It was just making my life slightly worse.
You could reasonably ask: why keep going there, then?
Because I didn't want to leave Substack. The medium was not the problem. I didn't need a new platform, or an elaborate alternate publishing system. I needed one stupidly specific thing: hide the feed when I didn't want to receive the full assault of everyone's opinions, while leaving the rest of Substack accessible.
Before AI, this annoyance would probably have remained an annoyance. It was too small to justify learning how to solve it. It was definitely too small to justify paying someone to solve it for me. Anyway, I don't think I would even have imagined it as a problem software could solve.
Worst-case scenario, I would eventually have stopped using a medium I liked because one part of its interface kept making me miserable.
Today, instead, I made a simple Substack feed blocker.
It did save me some irritation. It probably saved me a little time. But the biggest benefit was something else entirely:
I realized I no longer had to accept the interface as it was given to me.
Software is made for an imaginary user
Before AI, little tools on the internet broadly fell into two categories.
Sometimes, you found the exact thing you needed, made by some obscure developer you didn't know. You installed it without really understanding what was happening, crossing your fingers it wasn't malicious, and it solved your problem instantly. Wonderful.
More often, you found an elaborate piece of software that required a subscription and came with 200 functionalities, a dashboard and an endless onboarding sequence, when all you wanted was one tiny button.
That makes sense.
Commercial software has to solve sufficiently common and profitable problems to remain commercial software. A company cannot spend months building something for Marie specifically because Marie gets irritated when she sees a French man's hot take at the wrong time of day.
Tragic, but understandable.
The consequence is that almost every piece of software contains average assumptions about its users. It decides which frustrations matter, how work should be organized, what you probably want to do next and which compromises are considered reasonable. It also optimized for cash, and that tends to translate into more functions, for the people who are most willing to pay for them.
You probably don't resemble the imaginary user those decisions were made for.
So you adapt.
You develop workarounds. You tolerate the friction. You watch tutorials to learn the correct way to use a tool that was supposedly made to help you. Sometimes you even conclude that you are using it wrong, instead of considering that the software might simply be wrong for you. Maybe even sometimes, you develop a whole new obsession with a tool, buying into its cult-like following and you start to try to adapt more of your life to it (Notion (affiliate link), I'm looking at you).
The lesson underneath all this is that your frustration is not legitimate enough to organize the world around. Your preferences are whims. The software is Reality. It is your job to adapt to it.
Meh. I don't think that's particularly healthy.
And I think this is one of the most interesting things vibe coding changes for non-technical people. It's not that we can all become developers now. Nor is it that we can rebuild every platform for free (we can't).
But rather : we can suddenly find intensely specific solutions to intensely specific needs.
A preference can become a specification
You still need to understand what your problem actually is. You need enough technical literacy to know what a browser extension is, or what kind of file you are trying to create. You need a good dose of imagination and curiosity. You need to be willing to make mistakes. You need to test things and tell the AI when it has done something stupid.
But compared with learning an entire programming language before you can hide one feed?
It's night and day.
Once I understood that, I started noticing more of those silly, personal problems.
I was using Sublime to collect things I liked online. I'm a BIG believer in collecting as a creative practice. If you create, you need to magpie your way through the world. You collect colors for paintings, phrases for writing, situations for drawings, people who make you want to do better work.
That's how you make your taste tangible.
Pinterest is wonderful for discovery, but not always wonderful for sources. It is also deeply afraid of the tiniest nipple in a classical painting. I have spent more time than any reasonable person should arguing with Pinterest moderation that no, this 18th-century breast is not pornography.
Sublime was better suited to what I wanted. But it was still an external platform holding a collection I wanted to build over many years.
I wanted the original files. I wanted reliable sources. I wanted actual images, not just links that might die. I wanted my own posts and replies in there too. And I wanted AI to be able to work from that context.
So I built what I currently call my Taste Archive (see : i made my own thinking app).
In a few hours of talking to an AI, I had a browser extension that saved things into my own collection. It keeps the source. It keeps the local file when possible. It saves my Substack Notes, my replies, my tweets and whatever else I find interesting online.
It can tell if I'm trying to save my own Substack Note vs something I found while browsing. It has my collections, in the order I want them. It looks right on dark mode, I can hide the stuff I don't want to see right now...
It is a very bespoke solution, a silly personal whim. It may not matter to anyone but me.
But it works exactly as I want it to work.
Then there is Marie's Library.
Over the past few months, I've been making more HTML artefacts with AI: reports, analyses, interactive courses, little research decks and things I want to spend time with. HTML is great because it can be prettier and more usable than another wall of text in an md document.
The problem is that local HTML files mostly live on your computer.
I like reading and exploring things on my phone, in the evening, in bed, in ridiculous positions where a computer simply cannot enter the configuration. So I made myself a small password-protected website where I can publish those artefacts.
Now, when I ask an AI to create a report or analyze something, I can tell it to publish the result to my library. I don't necessarily read it immediately. I get to it later, when I actually want to be in that material.
Again: a stupidly personal preference.
In a conventional product brief, “Marie wants to read a pretty interactive report sideways in bed” would not be taken seriously. But as a personal software, it is enough of a reason to build the whole thing.
This is what I mean by permission software. Each tool gives practical legitimacy to something I would previously have treated as a whim. Together, they form a portrait of how I want to think, collect, learn and use the internet.
They adapt to me instead of asking me to adapt to them.
This isn't productivity
I want to be very clear about this: these tools have probably not saved me time.
At least not yet.
I've spent many hours building things that replace five-minute workarounds. Making something intensely personal requires you to understand your own preferences, explain them, test the result, change your mind and sometimes spend three hours fixing something you could have tolerated in existing software.
By conventional productivity maths, this is ridiculous.
But this is not conventional productivity. The value is elsewhere.
It's the recurring irritation you remove. It's what you learn about how systems work. It's the pleasure of an exact fit. It's the control you regain over your environment. It's the fact that the internet becomes a little nicer to inhabit when you can move the furniture around.
It's also how satisfying it is to notice a problem and make a real thing that solves it.
Will this tool repay the hours I spent building it? is not always the right question to ask. Sometimes it's : Does building this make my world a place I want to live in, even a little bit?
If it does, I think that counts.
Permission doesn't mean building everything
Now, because I have become the lady telling everyone she builds software with AI, people regularly tell me they want to replace their entire business platform with something they vibe coded themselves.
Payments. Authentication. Invoicing. Email. Course delivery. Video security. The whole thing.
And my answer is usually: eeeeh, be careful.
The fact that you can customize something does not mean you should.
I've worked with online entrepreneurs and course platforms for years. There are parts of a technical system that benefit from being intensely yours, and there are parts that benefit from being boring, reliable and maintained by someone else.
A personal archive benefits from specificity. A feed blocker benefits from specificity. My eccentric little library benefits from specificity.
Payments however... They don't need to express my personality. They need to work.
The distinction I'm arriving at is this: build where specificity creates value. Buy where reliability creates value. If making it yourself creates more responsibility, maintenance, and fragility than pleasure or control, leave it alone.
Your tools are evidence about you
Here is the slightly woo-woo part.
Building things for myself forced me to ask surprisingly intimate questions.
What do I really want? How do I actually work? What makes me want to return to a tool? What drains me? Which compromises have I accepted only because I assumed there was no alternative?
My small collection of personal software became not just a technical setup. But evidence of who I am and how I want my world to function.
My tools say I want access without the feed. I want connection without constant social performance. I want my collections to remain mine. I want the original files and the sources attached. I want AI to work from my context. I want information to be available where I will actually consume it, even if that means sideways in bed.
I already knew some of those things about myself. But translating them into specifications makes them concrete.
Which leads me to believe that personal software is a form of self-knowledge.
The interesting part is the permission though.
--
REVIEW OF THE AI EXPERIMENT :
This was interesting. I've edited this bit, but the cleaned up version made by Ai was pretty good. It helped that I had already made a skill that knows how I like to write, and that the transcript was EXTENSIVE. Actually, that's the part that I'm most impressed with. A 20 minute yapping session vs a writing one got me at least double the input. When I write, I tend to condense things, and not expand. If I had written this directly, it would have been a lot shorter. In fact, I kind of feel like I would have taken each part and made a post for each. This would have been 3 or 4 posts, rather than a single one.
I did cut A LOT from the end. A whole section and some more. It was dragging its argument. I'm tempted to also take the part about 'what should you build vs what you shouldn't' out too. Maybe it is it's own stuff. But I'll leave it here for the sake of the experiment.